a
https://doi.org/10.1038/s41698-025-01092-4
Multi-modal characterization of metabolic and immune gene clusters in adrenocortical carcinoma treatment
☒ Check for updates
Wenjun Hao1,2,3,4,9, Luhan Yao5,9, Yanlong Wang1,2,3,4,9, Jiayu Wan5, Yuyan Zhu6, Zhihong Dai1,2,3,4, Xu Sun7, Bo Fan1,2,3,4, Yuchao Wang1,2,3,4, Hao Xiang8, Xiang Gao1,2,3,4, Peng Liang1,2,3,4, Haolin Zhao 1,2,3,4, Liang Wang1,2,3,4, Ying Wang1,2,3,4, Hongyu Wang5, Deyong Yang8 X & Zhiyu Liu1,2,3,4 ☒ ☒ ☒
Adrenocortical carcinoma (ACC) is an uncommon and aggressive endocrine malignancy, characterized by limited therapeutic options and considerable variability in patient outcomes. The challenge is to combine the complex information of ACC with artificial intelligence (AI) and clinical and pathology data to achieve precision medicine and improve patient prognosis. We developed the Steroid-related Immune Score (SIS) using multi-modal analysis of genomics, digital pathology, and artificial intelligence and validated it in external datasets. In addition, we conducted single-cell RNA sequencing (scRNA- seq) of small samples and in vitro functional experiments. SIS delivered a stable performance with an AUC of 0.8 + 0.01 in the ResNet50 and Vision Transformer-B16 models. We validated the best model in external ACC cohorts. Using Class Activation Maps (CAMs) technology revealed that SIS was associated with lymphocyte infiltration, establishing it as a new feature in addition to the Weiss scoring system. Patients in the high SIS group responded well to immunotherapy, while the low SIS group showed adaptability to hormone inhibition therapy. Single-cell RNA sequencing data revealed the relationship between the tumor microenvironment and drug resistance in ACC. In vitro functional assays demonstrated that elevated DHCR7 gene expression correlated with unfavorable prognosis and treatment sensitivity, identifying it as a prospective therapeutic target. Furthermore, there are similarities between the metabolic characteristics of ACC and schizophrenia, such as calcium and iron ion levels. Our multi-modal analysis comprehensively characterizes the immune microenvironment of ACC, emphasizing the synergistic regulation of metabolic and immune gene clusters that influence ACC patients’ responses to immune and hormone therapies.
Adrenocortical carcinoma (ACC) is a rare and malignant endocrine tumor with an incidence of 0.5-2.0 cases per million people and it is more common in women1. The prognosis of patients is directly related to the stage of the tumor. The 5-year survival rate for stages I-III is over 50%. For stage IV, it is a mere 13%2. However, the overall 5-year survival rate is only 16-47% due to the heterogeneity and aggressiveness of ACC3. Despite surgery being the
main treatment and significantly improving survival, local or metastatic recurrence is common after surgery, often occurring within two years4,5. For patients with advanced and metastatic ACC, mitotane is currently the only FDA-approved first-line drug. However, its clinical efficacy is limited by the prolonged time required to reach therapeutic drug concentrations and severe adverse effects6. Existing treatment strategies are simply not enough
1Department of Urology, The Second Affiliated Hospital of Dalian Medical University, Dalian, Liaoning, China. 2Liaoning Provincial Key Laboratory of Urological Digital Precision Diagnosis and Treatment, Dalian, Liaoning, China. 3Liaoning Engineering Research Center of Integrated Precision Diagnosis and Treatment Technology for Urological Cancer, Dalian, Liaoning, China. 4Dalian Key Laboratory of Prostate Cancer Research, Dalian, Liaoning, China. 5School of Information and Communication Engineering, Dalian University of Technology, Dalian, Liaoning, China. 6Department of Urology, The First Affiliated Hospital of China Medical University, Shenyang, Liaoning, China. 7Department of Pathology, The First Affiliated Hospital of Dalian Medical University, Dalian, China. 8Department of Urology, The First Affiliated Hospital of Dalian Medical University, Dalian, Liaoning, China. 9These authors contributed
equally: Wenjun Hao, Luhan Yao, Yanlong Wang. ☒ lzydoct@163.com
e-mail: whyu@dlut.edu.cn; yangdeyong@dmu.edu.cn;
中 THE HORMEL INSTITUTE UNIVERSITY OF MINNESOTA
to meet the clinical needs of ACC patients. This situation underscores the pressing need for more precise molecular subtyping and targeted ther- apeutic strategies to optimize personalized treatment and improve prognosis.
In recent years, with the development of precision medicine, pathology analysis has gradually evolved from traditional morphological observation to digitalization and intelligence. The application of high-resolution whole- slide images (WSIs) has greatly improved the acquisition and storage effi- ciency of pathology data, laying the foundation for the application of arti- ficial intelligence (AI) in pathology7. AI-based pathology analysis is capable of mining new imaging features from large-scale data and integrating them with molecular histology data to improve the precision of tumor typing. However, existing AI pathology analysis is still limited to a single modality and fails to fully integrate information at the molecular level. Meanwhile, although genomics studies have revealed some of the driving molecules of ACC, their understanding of the tumor microenvironment (TIME) and drug resistance mechanisms is still incomplete due to the difficulty in resolving cellular heterogeneity with bulk RNA sequencing methods. Therefore, it is difficult for a single pathological analysis or molecular genomics study to comprehensively portray the molecular features and biological behaviors of ACC, limiting the application of precision medicine in this field.
This study transcends the traditional boundaries between pathology and molecular biology by, for the first time, employing an AI-driven pathological analysis and multi-omics integration approach to system- atically classify ACC. Through multimodal fusion, we not only refined the molecular typing of ACC based on pathological features but also analyzed the cellular heterogeneity of ACC using single-cell sequencing (scRNA-seq), overcoming the limitations of bulk sequencing methods. Additionally, we established a tumor immune microenvironment-based ACC classification system, uncovering the molecular mechanisms, drug resistance character- istics, and potential therapeutic targets of different subtypes. Furthermore, by integrating multi-dimensional drug prediction strategies, we identified candidate drugs tailored to distinct ACC subtypes, providing new avenues for precision therapy. Overall, this study represents a breakthrough in data integration, methodological innovation, and clinical translation, laying the groundwork for personalized treatment and novel target discovery in ACC.
Results
Establishment of the TIME subtype
According to recent literature, knowledge of the TIME should be the main emphasis of further ACC research8. Therefore, we analyzed 24 micro- environmental cell subpopulations using GSEA and classified ACC patients (TCGA + GSE76019 + GSE76021) into three clusters based on ssGSEA (Fig. la-c). The results were clear: TIME cluster A had the lowest ssGSEA scores, while TIME cluster C had the highest. In 23 microenvironmental cell subpopulations (excluding plasma cells), the three clusters exhibited sta- tistically significant differences in scores. Patients in TIME cluster B had the best prognosis (Fig. 1d). Furthermore, four key immune checkpoints (PD-1, PD-L1, PD-L2, CTLA4) exhibited the lowest expression in TIME cluster A (Fig. le-h).
Establishment of subtypes based on DEGs
Our analysis of the differential genes between the TIME clusters identified 18 common DEGs (Fig. 2a). These 18 genes underwent unsupervised hierarchical clustering analysis to find that the clustering stability was optimum at k = 2 (Fig. 2b, Supplementary Fig. 2a-c). We validated the classification by repeating the clustering analysis on the original dataset and two independent GEO datasets (GSE33371 and GSE10927). The results align with the previous classification pattern (Fig. 2c and Supplementary Fig. 2d-f). We named these groups Gene clusters A and B and plotted the expression heatmap of the 18 differentially expressed genes (Fig. 2e). Single- cell RNA sequencing data showed that these 18 genes were highly expressed in specific cell subsets and associated with immunity and metabolism (Fig. 2f). Patients in the Gene cluster B group had a better prognosis (Fig. 2g).
This group showed high infiltration abundance in 19 microenvironment cell subsets and significantly high expression of four key immune checkpoints (Fig. 2h, i). Finally, we calculated the Steroid-related Immune Score (SIS) by PCA. We separated the 142 ACC patients into groups with high and low SIS based on the optimal cutoff value (1.041494).
AI validation of SIS and associated pathological features
Deep neural networks are an indispensable tool for medical image analysis. They can identify features that are difficult to detect by the naked eye, such as microsatellite instability (MSI), tumor mutation burden (TMB), and gene expression status9,10. This study used deep learning technology to analyze whole slide images (WSIs), verified the effectiveness of SIS grouping, and explored its pathological characteristics to understand better the uniqueness of ACC (Fig. 3a-d). We used two mainstream deep learning networks, ResNet50 and Vision Transformer-B16, for validation and comparison (Table 1). The five-fold cross-validation results demonstrated that the AUC of the SIS classification reached 0.8 ± 0.01, with ResNet50 performing best (AUC=0.8214, accuracy =0.71) (Fig. 3e, f). Furthermore, ResNet50 excelled in the binary classification C1A/B model prediction of ACC (AUC=0.848, accuracy = 0.74) (Supplementary Fig. 3b, c), confirming the efficacy of SIS grouping in pathology. We created a heatmap in the original slice (Fig. 3g) to more clearly visualize the SIS grouping and its classification probability. We used Class Activation Maps (CAMs) technology to visualize the pathological features associated with the SIS subgroups. The findings indicated that the SIS correlated with sinusoidal invasion and necrosis in the Weiss score, and surprise, with lymphocytic infiltration (Fig. 3h and Sup- plementary Fig. 3d). To enhance the verification of the model’s clinical application, we picked the ResNet50 model exhibiting optimal performance. We validated ACC patients from the First Affiliated Hospital of Dalian Medical University and the Second Affiliated Hospital of Dalian Medical University. The findings indicated that patients in the high SIS group exhibited significant lymphocytic infiltration. Conversely, patients in the low SIS group had minimal lymphocytic infiltration, with both groups marked by sinusoidal invasion and necrosis (Fig. 3i, Supplementary Fig. 3e). This suggests that the model exhibited consistent performance in the external validation set and possesses potential therapeutic applications. Subsequent prognostic follow-up assessments indicated that patients in the high SIS group had significantly prolonged survival periods compared to those in the low SIS group. However, owing to the restricted follow-up duration and patient population, the p-value failed to achieve statistical significance (p > 0.05). Survival studies, after integrating these patients with those from the TCGA cohort, indicated a decrease in the p-value, implying that these patients align with the prognostic characteristics of the SIS sub- group (Supplementary Fig. 3f, g, Supplementary Fig. 4a).
The new ACC subgroup SIS closely related to clinical outcomes
Several subtypes have been identified in the ACC dataset, including the COC and C1A/C1B subtypes(Fig. 4a)11,12. The high SIS group was more likely to correlate with the positive prognosis COC1 and the inactive C1B subtype. Conversely, the low SIS group was significantly correlated with the unfavorable prognosis of COC2 and COC3, as well as the aggressive C1A subtype. There are significant differences in SIS between the COC subtype, C1A/C1B subtype, expression subtype, and methylation subtype (Fig. 4b, c and Supplementary Fig. 4d, e). The TCGA immune subtype analysis shows that ACC patients are mainly distributed in the C3 group with the best prognosis and the C4 group with the worst prognosis. The C3 group had the highest number of high SIS patients, while the C4 group had the highest number of low SIS patients. Patients in the C3 group had a SIS level much greater than those in the C4 group (Fig. 4d). Patients with high SIS have a favorable prognosis in the TCGA and GEO datasets (Fig. 4e, f, and Sup- plementary Fig. 4a-c). Univariate and multivariate Cox regression analysis clearly identified SIS as an independent prognostic factor for ACC patients (Fig. 4h, i). The low SIS group was composed mainly of patients with the following clinical parameters: female gender, T4, N1, M1, Stage III, and Stage IV (Fig. 4j). The Weiss score is crucial in the pathological diagnosis of
a
b
TIME cluster
T cells CD4 memory activated
Mast cells activated
B cells naive
TIME cluster
3
100
T cells CD4 naive
A
Adaptive and activated innate immune cells
Plasma cells
2
B
C
B cells memory
C
T cells regulatory Tregs
T cells gamma delta
1
T cells CD4 memory resting
TIME cluster
0
NK cells activated
50
Dendritic cells activated
-1
Macrophages M1
PC2
A
T cells CD8
-2
B
T cells follicular helper
c
NK cells resting
-3
Inactivated Innate
Mast cells resting
0-
immune cells
Dendritic cells resting
Macrophages M2
Monocytes
Eosinophils
Macrophages MO
Other stromal
Neutrophils
cells
Endothelial cells
-50
Fibroblasts
-40
0
40
PC1
C
d
TIMEcluster E
1 Z
B B
C
1.00-
TIME cluster
A
1.00-
**
·
**
ns
*
**
**
8
0.75
+ B
+ c
SSGSEA score
Survival probability
0.75
·
0.50
0.50
D
-
5
.
C
T
0.25
·
O
1
0
A
T
0.25
6.236e-04
0.00
B cells memory
B cells naive
Dendritic cells activated
Dendritic cells resting
Endothelial cells
Eosinophils
Fibroblasts
Macrophages MO
Macrophages M1
Macrophages M2
Mast_cells_activated
Mast cells resting
Monocytes
Neutrophils
NK cells activated
NK cells resting
Plasma cells
T cells CD4 memory activated
T cells CD4 memory resting
T cells CD4 naive
T cells CD8
T cells follicular helper
T cells gamma delta
T cells regulatory Tregs
0.00
0
1
00
A
Un
1
8
a
10
11
12
13
14
15
16
TIME cluster
Time(years)
Number at risk
QUE
6
0
14
ON
DON
DON
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Time(years)
e
f
g
h
TIME cluster
A
B
c
TIME cluster
A
B
c
TIME cluster
A
B
c
TIME cluster
A
B
C
ns
4
6
5
N
4
.
ns
5
ns
!!
.**
4.
3
CTLA4 expression
1
PD-L1 expression
PD-L2 expression
4
3
PD-1 expression
3
2.
2
2.
T
2
I
1
1
1
:
0
0
A
B
C
À
B
Č
A
B
C
A
B
C
ACC. The low SIS group accounted for over 50% of the Weiss score-related indicators, including necrosis, mitotic rate > 5/50 HPF, venous invasion, and sinusoidal (lymphatic) invasion (Fig. 4g). Epithelial-to-mesenchymal transition (EMT), tumor mutational burden (TMB), mRNA expression- based stemness index (mRNAsi), and Ki-67 are common biomarkers in oncology. ACC patients with high EMT, high TMB, high mRNAsi, and high Ki-67 generally have a poor prognosis, while patients in the low SIS group within these stratifications have the worst survival outcomes and vice versa (Fig. 4k-n). We identified mutations in known ACC genes (e.g., CTNNB1, ZNRF3)12 and also found mutations in some new genes (e.g., TTN, HLTF, ADAMTS16, NSD1, and PARP8) (Fig. 40). The only exception was HLTF, which exhibited a high mutation rate in the high SIS group, while the other genes were almost exclusively mutated in the low SIS group.
The low SIS group closely related to steroid synthesis
According to KEGG GSEA and GSVA analysis, the high SIS group exhibited enrichment in immune-related pathways, whereas the low SIS group had enrichment in steroid biosynthesis pathways. The HALL- MARK GSEA and GSVA analyses indicated that the high SIS group had
enrichment in immunological pathways, while the low SIS group demonstrated enrichment in cholesterol homeostasis pathways (Fig. 5a, b, Supplementary Fig. 6a, b). We speculate that patients with high SIS may be related to immune response, while patients with low SIS may be related to adrenal function. SIS was substantially negatively correlated with the adrenal cortical differentiation index (ADS) (r = - 0.62) (Fig. 5e). Most patients with high SIS were predicted to have low ADS (71.4%) and did not have abnormal hormone (61.5%) or cortisol (80.8%) secretion. In con- trast, most patients with low SIS were predicted to have high ADS (66%) and had abnormal hormone (78.7%) and cortisol (57.4%) secretion (Fig. 5d). In the cortisol present group, 84% of patients exhibited low SIS or high ADS. In the hormone present group, 79% of patients had low SIS, which is greater than the 70% of patients with high ADS. Therefore, low SIS is a more effective assessment of adrenal function than high ADS (Fig. 5f, g). We screened nine core genes related to steroid hormone synthesis by HALLMARK_CHOLESTEROL_HOMEOSTASIS and KEGG _- STEROID_BIOSYNTHESIS analysis (Fig. 5c). Mitotane is the main drug to inhibit ACC steroid hormone synthesis, which mainly inhibits SOAT1, which is related to cholesterol storage, and CYP11A1 and CYP11B1,
a
b
C
d
B-A
C-A
consensus matrix k-2
consensus matrix k=2
100
51
0
O
V
47
400
0
0
50
PC2
Gene cluster
18
A
19
614
B
.
B
0
A
219
-50
-40
0
40
C-B
PC1
e
f
g
4 TumorfPurity
0.9
1.00
CYP17A1
TumorPurity ESTIMATEScore
0.3
Gene cluster
2
ESTIMATEScore
ABAT
1.00
Immune Score
StromalScore
4000
CD52
0.80
A
Stage
0.75
B
1
Fustat
0
Project
-2000
CORO1A
0.60
TIMEcluster geneCluster
ImmuneScore 2000
CXCL12
0.40
Survival probability
AF1
2
HCLS1
CIDA
1000
PLEK
0.20
0.50
C108
StromalScore
C1QB
0.00
CD163
2000
SIGLEC1
CD52
COROTA
1000
MS4A4A
CXCL10
Stage
0.25
Stage I !!
F13A1
p=0.003
CXCL12
Suge Ill Stage IV
CD163
CXCLD
L
AIF1
F1341
Fustat
0.00
FCERIG
C1QA
Dead
FOUR2
Project
CXCL9
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
HCLS1
GEO
FOLR2
Time(years)
MSAA4A
TOGA
CXCL10
PLEK
TIMEcluster
Number at risk
SIGLECI
8
FCER1G
C
AMAT
Macrophage
T_cell
Cancer_cell
Progenitor_cell
Endothelial_cell
Mesenchymal_cell
Gene cluster
10690 67 49 37 28 17 14 13 11 7 5 4 2 2 1 1
CYP17A1
Genecluster
B
B
36 35
26
23 19
13
10
8
5
3
2
2
2
1
1
1
1
0
1
2
3
4
5
6
7
8
9
10
12
1 13
14
5
16
1
geneType
B
Time(years)
h
Gene cluster
A
B
Gene cluster
A
B
Gene cluster
A
B
4.
5-
1.00
*
ns
**
*
**
ns
ns
**
4
PD-L2 expression
3
PD-1 expression
0.75
&
:
2
ssGSEA score
:
2
0.50
…
&
.
1
1
:
S
A
B
A
8
0.25
.
Gene cluster
A
B
Gene cluster
A
B
3.5
0.00
5
-
-
B cells memory
B cells naive
Dendritic cells activated
Dendritic cells resting
Endothelial cells
Eosinophils
Fibroblasts
Macrophages MO
Macrophages M1
Macrophages M2
Mast cells activated
Mast cells resting
Monocytes
Neutrophils
NK cells activated
NK cells resting
Plasma cells
T cells CD4 memory activated
T cells CD4 memory resting
T cells CD4 naive
T cells CD8
T cells follicular helper
T cells gamma delta
T cells regulatory Tregs
3.0
4
CTLA4 expression
2.5
PD-L1 expression
3
2.0
2
1.5
1.0
1
À
B
A
B
which are associated with the conversion of cholesterol to cortisol and aldosterone (Fig. 5m). Among the 12 core genes, DHCR7, CYP51A1, SOAT1, and CYP11A1 were highly correlated with SIS and ADS. DHCR7 was significantly and stably differentially expressed among them in the hormone-related subgroup, while the other genes were more variable (Fig. 5h). Except for EBP, CYP11A1, and CYP11B1, genomic alterations were more common in the low SIS group compared to the high SIS group, particularly for DHCR7, SOAT1, and FDFT1, which showed no altera- tions in the high SIS group but ≥10% alterations in the low SIS group (Supplementary Fig. 5a). Among these 12 genes, only patients with
DHCR7 genomic alterations showed poor prognosis in overall survival (OS), progression-free survival (PFS), and disease-specific survival (DSS) (Supplementary Fig. 5b-d). In contrast, patients with SOAT1 genomic alterations showed poor prognosis only in PFS (Supplementary Fig. 5e-g). The metabolic genes most associated with SIS features, DHCR7 and SC5D, were further investigated by Lasso regression, Random Forest(RF), and Support Vector Machine - Recursive Feature Elimination(SVM-RFE) machine learning methods (Fig. 5i). In the pan-cancer expression analysis, DHCR7 showed the highest expression in ACC, while SC5D exhibited an intermediate expression level. Nevertheless, high expression of both was
a
b
C
e
AUCs of ResNet50
1.0
0.8
0,6
È
0.4
cross 1 (AUC = 0,786)
cross 2 (AUC = 0,893)
cross 3 (AUC = 0.714)
0.2
cross 4 (AUC - 0.893)
cross 5 (AUC = 0.821)
Mean AUC (AUC = 0.819 ± 0.068)
0.0
± 1 std, dev,
0.0
0.2
0.4
0.6
0.8
1.0
d
f
FPR
AUCs of Vision Transformer
1,0
0.8
0.6
0.4
cross 1 (AUC = 0.679)
cross 2 (AUC = 0.786)
cross 3 (AUC = 0.893)
0,2
cross 4 (AUC = 1,000)
cross 5 (AUC = 0,607)
Mean AUC (AUC = 0.790 ± 0.142)
0,0
± 1 std. dev.
0,0
0.2
0.4
0.6
0.8
1.0
FPR
g
h
Input
High SIS probability
0
0
50
50
300
100
150
150
high
200
200
250
250
9
0
50
100
250
200
250
0
50
300
150
200
250
0
low
S
9
S
50
50
aDO
100
150
150
200
200
250
0
50
100
150
200
250
250
0
50
200
150
200
250
Input
Low SIS probability
High SIS
Low SIS
i
0
0
50
50
100
100
150
150
200
200
250
250
D
50
100
150
200
250
0
50
200
150
200
250
0
.
0
50
50
100
100
150
150
200
200
250
D
50
100
150
200
250
250
0
50
100
150
200
250
High SIS
Low SIS
associated with poorer patient prognosis (Fig. 5j and Supplementary Fig. 5h-j). After extensive analysis, we concluded that DHCR7 is a potential molecular marker for ACC. Single-cell RNA sequencing results showed that DHCR7 was specifically expressed in cancer cells (Fig. 5l).IHC results showed elevated DHCR7 protein expression in tumor tissues (Fig. 5k). CCK-8 assay results showed that NCI-H295R cells (IC50 = 4.684 uM) were more sensitive to Mitotane than SW-13 cells (IC50 = 6.918 µM)
(Fig. 5l, m). Mitotane sensitivity was significantly increased in both DHCR7 knockdown cells (Fig. 5n, o).
The high SIS group closely related to immunity
The findings from GSEA and GSVA indicated that patients in the high SIS group exhibited more pronounced tumor immune features Supplementary Fig. 6a, b). Therefore, we investigated the relationship of SIS with the
| Model | AUC | Accuracy | Recall | Precision |
|---|---|---|---|---|
| SIS Vision Transformer-B16 | 0.793 | 0.7452 | 0.6336 | 0.65 |
| Resnet50 | 0.8214 | 0.7088 | 0.65 | 0.65 |
| C1A/B Vision Transformer-B16 | 0.732 | 0.6233 | 0.62 | 0.54 |
| Resnet50 | 0.848 | 0.7367 | 0.62 | 0.79 |
ESTIMATE score and the abundance of immune cell infiltration more closely. The results showed that the high SIS group had a much stronger correlation with the ESTIMATE score than the low SIS group did within the same dataset. In fact, the high SIS group had a correlation that was above 0.8 in both datasets (Fig. 6a). Regarding the abundance of immune cell infil- tration, both groups showed correlations with increased monocytes, decreased neutrophils, decreased basophils, decreased natural killer cells, decreased naive CD8 T cells, and increased cytotoxic cells. The abundance of CD8 T cells and Tfh cells in the high SIS group increased with increasing SIS.
a
b
C
d
SIS
COC E COCI E coCZ E coc3
CIA/C1B ẸP CIA E C1B
Immune Subtype 9 ca # C4
COC
0.017
5.9e-09
0.00015
7.10-07
C1A/C1B
0.041
10
Immune
10
5
ª
Expression
5
9
35
2
Methylation
-
0
0
Weiss Score
0
.
10
Expression (K mal)
DOCT COC2 INI COCa MI NA
U
Steroid-pheno-high
A
-5
-5
CAD
COC1
COC2
COC3
CIA
CİB
C3
C4
e
f
g
1.00
1.00
Weiss scoring system
Absent (0)
Present(1)
SIS
SIS
Clear cells s 25% of the tumor volume
17%
83%
High
High
Survival probability
0.75
-
Low
0.75
Low
High Fuhrman nuclear grade (Ill or IV)
21%
Survival probability
79%
Necrosis
23%
77%
0.50
0.60
Diffuse architecture > 30% of tumor volume
31%
69%
Mitotic Rate > 5/50 HPF
41%
59%
-
0.25
Capsular invasion
42%
55%
2.281e-04
0.25
2.498e-94
Atypical mitosis
48%
52%
0.00
0.00
Venous invasion
56%
44%
0
1
2
3
4
5
6
8
9
10
11
12
13
14
15
16
0
1
2
3
4
5
6
8
9
10
11
12
13
14
15
18
Time(years)
Time(years)
Sinusoidal (lymphatic) invasion
58%
42%
**
Number at SIS
Number at SIS
100
50
0
50
Percentage
100
SIS
High
52
48
37
31 26
21
16
13
10
8
4
3
3
2
2
1
1
SIS
ligh
67
59
46
38
33
25
18
15
11
B
4
3
3
2
Z
1
1
90
77
56
41 30
20
11
9
B
6
5
4
3
1
1
1
OW
122
85 1
66
46
35
25
16
13
12
&
&
V
6
2
Z
2
1
0
1
2
3
4
5
6
G
8
9
10
11
12
13
14
15
16
0
2
3
4
5
6
7
8
9
10
11
12
13
14
15 16
SIS
high
low
high
low
Time(years)
Time(years)
h
pvalue
Hazard ratio
pvalue
Hazard ratio
,
Age
0.328
1.012(0.988-1.038)
Gender
0.972
0.986(0.451-2.154)
T
0.007
2.837(1.335-6.030)
22%
22%
I
T
<0.001
3.378(2.110-5.407)
87%
80%
62%
82%
67%
50%
57%
89%
62%
M
0.403
1.777(0.462-6.840)
83%
80%
81%
80%
N
0.152
2.038(0.769-5.400)
SIS
100%
-
M
<0.001 6.150(2.710-13.959)
Stage
D.761
0.864(0.336-2.221)
78%
70%
1
Stage
<0.001
2.658(1.714-4.121)
49%
48%
53%
31%
E0%
38%
43/5%
42%
45%
SIS
0.003
0.193(0.066-0.566)
4
SIS
0.038
0.303(0.099-0.939)
17%
EOW
19%
20%
0
2
4
6
8
10
12
0
1
2
3
4
5
6
-
[n= 27)
0 = 500
|= = 290
(n = 42)
(n= 8)
|= = 68)
(n = 9) (n = 82)
n = 15)
n= 37)
(n = 16)
Hazard ratio
Hazard ratio
mais
não
PENALE
MALE
Ti
12
T
14
Type
NO
NI
M
Bhaçel
Bhaçılı
6hçelIl
9gs I
k
m
n
O
VON
120
4.75
L-K-ST
Sunivel probability
H-EMIT L-EMT
an
A
0,75
H-TMOD L-TMg
Survivalprobably
0.70
p value
0.50
ZNRF3
34
3.57
1.8020-3
p=0.001
LE
p=0.014
p<0.001
QUES
P=0.001
CTNNB1
26
2.902e-3
£
10 11 12
ILO
1
£
10 11 12
9
·
4
1
10 11 to
Number at rid
Number at risk
Number all risk
Number at risk
TTN
18
0.0226
0
2
-SANA
3
te
10 11 T
0
SIS
HLTF
10.71
0.0431
High
Low
wDC
1.
ADAMTS16
14
0.0454
·
H-EMT-H-GE
1.
Survival probably
- 1-10-47-1-88
4
- L-EMT.L-818
Survival probubily
ov
L-TVA+L-GE
HARNAS-L-88 - L-ORNAR4H-SIS
-
NSD1
14
.
0.0454
05
:
Su
PARP8
14
ـى
P=0.001
02
0.0454
p<0.001
NE
0-0. 001
SA
OU
VI
0
10
20
30
4
# 10 11 12
6
Alteration event frequency (%)
Number at tisk
Number at risk
Nurbur at risk
Number at disk
O
8
S
1
—!
GSE10927) (log-rank test, P = 2.498e-04). g The distribution of SIS subgroups in the Weiss Scoring System is shown in a Likert plot. * P < 0.05, ** P < 0.01, *** P < 0.001. h, i Univariate and multivariate analysis of clinical characteristics. j Distribution of SIS subgroups across clinical characteristics. k-n Kaplan-Meier curves for the pre- diction of overall survival in ACC patients with Ki-67, EMT, TMB, and mRNAsi and their stratification in SIS. o Genomic alterations in SIS subgroups.
a
b
C
HALLMARK_CHOLESTEROL_HOMEOSTASIS
KEGG_ANTIGEN_PROCESSING_AND_PRESENTATION
KEGG_B_CELL_RECEPTOR_SIGNALING_PATHWAY
- HALLMARK_COMPLEMENT
0.
KEGG_CELL_ADHESION_MOLECULES_CAMS
HALLMARK_IL2_STATS_SIGNALING
HALLMARK_JILO_JAK_STATS_SIGNALING
KEGG_CYTOKINE_CYTOKINE_RECEPTOR_INTERACTION
0.4
Enrichment Score
KEOG_CYTOSOLIC_DNA_SENSING_PATHWAY
HALLMARK_INFLAMMATORY_RESPONSE
KEOG_FC_EPSILON_RI_SIGNALING_PATHWAY
HALLMARK_INTERFERON_ALPHA_RESPONSE
KEGG_NATURAL KILLER_CELL_MEDIATED_CYTOTOXICITY
Enrichment Score
- HALLMARK_INTERFERON_GAMMA_RESPONSE
9.
KEGG_T_CELL_RECEPTOR_SIGNALING_PATHWAY
- KEGG_TOLL_LIKE_RECEPTOR_SIGNALING_PATHWAY
CL
17
9
4
-0.4
- KEGG_CELL_CYCLE
- KEGG_NUCLEOTIDE_EXCISION_REPAIR KEGG_PROTEIN_EXPORT
- HALLMARK_MYC_TARGETS_V1
- HALLMARK_CHOLESTEROL_HOMEOSTASIS
-0.4
- KEGG_STEROID_BIOSYNTHESIS
- HALLMARK_O2M_CHECKPOINT
-0.8
KEGG_STEROID_BIOSYNTHESIS
High SiSe
>Low SIS
High Sis<
Low SIS
d
h
SIS
Correlation
ADS
0
1
LogFC
Hormone
-1.50
0
Cortisol
Hormone
Cortisol
SIS ☐ High
Low
ADS ☐ High
☐ Low
Hormone ☐ Present
☐ Absent ☐ NA
Cortisol
☐ Present ☐ Absent
NA
SIS
ADS
SIS
ADS
e
f
g
DHCR7
-0.61 0.72
-1.17-1.30-0.82-0.71
SOAT1
R= 0.02. p=10-09
100% -
p =0.02
p = 0.456-05
p=0.24
p =8.200-05
p =0.88
p = 1.01e-04
100% -
p=0.02
p =0.02
P =0.04e-03 p =5.58e-03 p =7.0Je-03 p =1.05e-04
0.55 0.57
-0.89-0.60 -0.49-0.45
CYP51A1
0.52 0.58
-0.94 -1.02-0.76-0.80
1
90% -
.
90% -
29%
23%
29%
CYP11A1
-0.94 -1.25-0.86
80% -
46%
38%
49%
80% -
0.51 0.81
D
70% -
70%-
SQLE
66%
0.47 0.54
-1.08-1.34 -0.81-0.90
60% -
80%
79%
84%
SIS
60% -
70%
ADS
-0.59 -0.91 -0.66-0.51
2.
84%
SC5D
0.46 0.62
50% -
low
50% -
high
FDFT1
0.45 0.59
-0.82-1.02-0.60-0.56
40% -
high
40% -
low
HSD17B7
0.44 0.61
-0.59-0.77 -0.54-0.53
.
30% -
54%
62%
71%
77%
71%
30% -
EBP
0.44 0.65
-0.60 -1.01 -0.65-0.58
20% -
51%
20% -
20%
21%
34%
30%
LSS
10% -
-0.39 0.61
-0.74 -1.01 -0.58-0.62
-3
10% -
16%
16%
NSDHL
0% -
-0.39 0.63
·
-0.42-0.96-0.46-0.53
90
(n = 41) high ADS
6
(n = 37)
(n = 26) Hormone Absent
(n = 47) Hormone Present
(n =41) Cortisol Absent
(n = 32)
0%-
S
(n = 28)
(n = 50)
(n = 26)
(n = 47)
(n = 41)
(n = 32)
0.38
SIS
5
low ADS
Cortisol Present
high SIS
low SIS
Hormone Absent
Hormone Present
Cortisol
Absent
Cortisol Present
CYP11B1
İ
j
k
Expression Level
Lasso
RF
DHCR7
High
+
DHCR7
Low
3
0
6
1
1.00-
2
Survival probability
0.75
DHCR7
1
2
0
0
0
0.50
Macrophage
T_cell
Endothelial_cell
Mesenchymal_cell
Cancer_cell
Progenitor_cell
0.25
p<0.001
0
0.00
0
1
2
3
+
5
6
7
8
9
Time(years)
10 11 12
Tumor
Normal
SVM
m
n
SW-13
O
NCI-H295R
120
120
100
IC 50= 6.918 μ. Μ
100
IC 50= 4.684 μ. Μ
Cell viability%
80-
Cell viability%
80-
Cortisol
60
60
40
40
Corticosterone
Aldosterone
Lipid droplet
CYP11B1
Mitotane
20
20
StAR
cholesteryl esters
0
0
Pregnenolone
.
CYP11A1
cholesterol
+
-20
5
10
15
20
25
-20
2
4
6
8
10
12
Mitotane(uM)
Mitotane(uM)
SOATT
free cholesterol
Endoplasmic reticulum
1.5-
1.2-
si NC
si DHCR7-1#
1.5-
1.2-
si NC
si DHCR7-1#
DHCR7
Relative DHCR7 mRNA
si DHCR7-2#
expression
cell viability
Relative DHCR7 mRNA
si DHCR7-2#
7-dehydrocholesterol
1.0-
1.0-
1.0-
0.8-
expression
1.0-
cell viability
0.8-
0.6-
0.6-
0.5
0.4-
0.5-
0.4-
0.2-
0.2-
0.0
si NC
0.0
0.0
si DHCR7-1#
si DHCR7-2#
Control
Mitotane
0.0
si NC
si DHCR7-1#
si DHCR7-2#
Control
Mitotane
Nucleus
In contrast, in the low SIS group, the abundance of macrophages, macro- phages M1, macrophages M2, and activated NK cells increased with lower SIS (Fig. 6b). Through machine learning analysis, pan-cancer immune studies revealed four immune-related factors-MHC, CP, EC, and SC13. ACC analysis revealed higher MHC and EC levels, and lower CP and SC levels in the high SIS group (Fig. 6c). Meanwhile, the high SIS group exhibited a significantly higher immunophenoscore (IPS) (Fig. 6d). TIP analysis revealed higher activity scores for CD8 T cells, macrophages, NK
cells, and infiltration of immune cells into tumors (step 5) in the high SIS group (Fig. 6e). And the overall activity score strongly correlated with SIS (r = 0.69) (Fig. 6f). Analysis of the 5 immune expression signatures revealed higher, statistically significant scores in the high SIS group (Fig. 6g). In addition, immune infiltration is associated with DNA damage14. Our ana- lysis revealed lower SNV neoantigens, nonsilent mutation rate, copy number variation (CNV) burden (number of segments), and homologous recombination deficiency (HRD) in the high SIS group, in addition to the
Fig. 5 | SIS is associated with steroid hormone synthesis. a, b KEGG and HALLMARK GSEA in SIS subgroups. c Common genes screened from the core genes of the two enrichment pathways. d The distribution of the ensemble of ADS, hormone, and cortisol levels among SIS subgroups. e Correlation between SIS and ADS. f, g SIS and ADS are distributed over hormones and cortisol, respectively. h Correlation of steroid hormone-related genes and Mitotane primary target genes with SIS and ADS expression. Expression values were compared with low SIS, high ADS, hormone present, and cortisol present. i Lasso regression, RF, and SVM-RFE machine learning jointly screened for genes most associated with SIS. j Kaplan- Meier curve was used to predict DHCR7 overall survival. k Protein expression of DHCR7 in ACC tissues and normal adrenal tissues by IHC. l Distribution of DHCR7
expression in single-cell RNA sequencing. m Mechanistic pathways associated with DHCR7 and Mitotane. n Cell viability was assessed after treatment with different concentrations of Mitotane in SW-13 cells. Control or DHCR7 siRNA was trans- fected, incubated for 48 h, and then collected for RT-PCR analysis of the DHCR7 gene. Cell viability was assessed after treatment of SW-13 cells with control or DHCR7 siRNA transfected with Mitotane (6.9 uM) for 48 h. o Cell viability was evaluated in NCI-H295R cells treated with varying concentrations of Mitotane, as well as in cells transfected with control or DHCR7 siRNA and incubated for 48 h, followed by RT-PCR analysis of DHCR7 expression. Cell viability was assessed after treatment of NCI-H295R cells with control, or DHCR7 siRNA transfected with Mitotane (4.7 µM) for 48 hours. * P < 0.05, ** P < 0.01, *** P < 0.001.
group having a low proliferation rate (Supplementary Fig. 6f). Immunity quantitative trait loci (immunQTLs) were applied to assess the impact of genetic variants on immune infiltration. Survival-associated immunQTLs (FDR <0.05) GWAS analyses revealed the highest number of QTLs for regulatory T cells (Tregs), followed by CD8 T cells, and the lowest number for resting CD4 memory T cells (Supplementary Fig. 6g). Among these GWAS disorders, the highest number of calcium level-related QTLs were all from Tregs. Schizophrenia involves T-cell regulatory (Tregs), T-cell CD8, macrophages M1, and T-cell CD4 memory resting, with the widest cover- age. TIDE is an important indicator for predicting tumor response to immune checkpoint blockade (ICB). Both TCGA and GEO datasets showed that SIS was significantly negatively correlated with TIDE (Supplementary Fig. 6c, d). Over 70% of patients in the high SIS group responded to ICB, while more than 65% in the low SIS group were resistant. Notably, the low SIS group represented over 80% of ICB-resistant patients (Fig. 6h, i). TIDE’s two main immune escape mechanisms are associated with T cell dysfunc- tion and T cell exclusion, respectively15. Our data indicate that the immune escape mechanism in ACC is primarily associated with T cell dysfunction. Cancer-associated fibroblasts (CAFs), myeloid-derived suppressor cells (MDSCs), and the M2 subtype of tumor-associated macrophages (TAMs) are the three main cell types that suppress T-cell infiltration16. A significant negative correlation was observed between the high SIS group and TAM M2. In addition, the immune profile of ACC was strongly correlated with interferon-gamma (IFNG) response and T-cell inflammatory phenotype (Merck18)(Supplementary Fig. 6e)17. We further assessed the expression of three molecules associated with tumor escape mechanisms in the low SIS group-Most MHC molecules were underexpressed in the low SIS group, thus avoiding T cell recognition; immunosuppressive factors (e.g., TGFBR1) might be upregulated for tumor escape; and immunostimulatory factors (e.g., RAET1E) might be downregulated to avoid immune attack (Fig. 6j).
We selected a low-SIS patient (with abnormal hormone secretion) for single-cell RNA sequencing analysis to explore the interaction between the tumor and its microenvironment. The analysis revealed the cell sub- populations of this patient were mainly concentrated in cancer cells (78.95%), macrophages (15.44%), progenitor cells (3.42%), mesenchymal cells (1.17%), T cells (0.65%), and endothelial cells (0.38%) (Fig. 6k, Sup- plementary Fig. 7). We screened 79 exosome-associated genes specifically expressed in tumor cells (avg_log2FC>0.4), and KEGG showed that these genes were enriched in hormone synthesis and secretion pathways, in addition to metabolic reprogramming and energy metabolism pathways, immune evasion and TIME pathways, and drug metabolism and resistance pathways (Fig. 6l). Using HitPredict and CellphoneDB analyses, we found that these exosomes interacted with ligand receptors of immune cells, with more than 65% of the interactions involving T cells and macrophages. Most of these exosomes were highly expressed in the TCGA and GEO datasets in low SIS patients (Fig. 6m). These genes were strongly associated with tumor, extracellular matrix, and immune processes (Fig. 6n). Unexpectedly, some genes (e.g., AXL, HGF, PDGFB, and PDGFRB) were associated with EGFR tyrosine kinase inhibitor resistance.
Drug sensitivity prediction based on SIS subgroups
We successfully predicted drug sensitivity for SIS subgroups using the GDSC1 and GDSC2 datasets. The results showed that patients in the
high SIS group were significantly more responsive to drugs targeting the PI3K/Akt/mTOR pathway (90%). Additionally, drugs targeting the following pathways showed greater sensitivity in this group: 54.5% on the Autophagy pathway, 40.9% on PI3K/Akt/mTOR pathway, and 18.2% on the Protein Tyrosine Kinase/RTK pathway. Furthermore, the GDSC1 and GDSC2 datasets showed that BI-2536 (PLK1 inhibitor) was more effective in the low SIS group based on the TCGA and GEO datasets (Fig. 7a). SPIED3 also predicted increased sensitivity to the MTOR inhibitor in the high SIS group, consistent with the GDSC results (Fig. 7b). CMap analysis, on the other hand, confidently identified several drugs with increased sensitivity in the low SIS group, including calmodulin antagonists18,19, dopamine receptor antagonists20, oxidos- qualene cyclase inhibitors21, serotonin receptor antagonists22, and sterol demethylase inhibitors23. All of these drugs inhibited steroid hormone synthesis, supporting our finding that the low SIS group exhibited adrenal function characteristics (Fig. 7c). SPIED3 and CMap together predicted four classes of drugs to be effective for patients in the low SIS group: acetylcholine receptor antagonist (mebeverine), dopamine receptor antagonist, BCR-ABL kinase inhibitor (imatinib), and nor- epinephrine reuptake inhibitor (maprotiline) (Fig. 7d). Notably, all dopamine receptor antagonists are used for the treatment of schizo- phrenia. Furthermore, single-cell RNA sequencing revealed that the tumor cell-specific genes identified were significantly enriched in the CALCIUM ION BINDING and IRON ION BINDING pathways, both of which are associated with schizophrenia, in the GO-MF enrichment analysis(Fig. 7e)24,25. The CHP1 gene plays a role in calcium metabolism and facilitates ferroptosis. The PCLO gene, linked to calcium metabo- lism, shows a significant association with schizophrenia(Fig. 7f)26.
Discussion
In this study, we definitively identified the metabolic and immune syner- gistic regulatory gene clusters through genomic analysis. We explored the molecular features of ACC in depth by combining digital pathology and AI technologies. We introduced a novel SIS-based molecular classification that uncovered the molecular heterogeneity of ACC and offered new insights for personalized treatment. Additionally, single-cell RNA sequencing revealed, for the first time, interactions between ACC tumors and immune cells. We discovered a new immune escape mechanism, which provides a theoretical basis for immunotherapy. Integrating multiple datasets and performing multidimensional analysis, we found that high SIS patients with better prognosis responded more favorably to ICB, while low SIS patients with poorer prognosis were more suited for hormone-based drug therapy.
The five-fold validation results of ResNet50 and Vision Transformer- B16 prove that the pathology images grouped by SIS exhibit high AUC values (0.8 ± 0.01). ResNet50 performs optimally, with an AUC value of 0.82 and an accuracy of 0.71. In addition, the Weiss score is an essential indicator for ACC diagnosis, and our study revealed the association between SIS subgroups and sinusoidal invasion and necrosis. Meanwhile, we unex- pectedly found the correlation between SIS subgroups and lymphocytic infiltration, which was also confirmed in the external validation set. This finding implies that SIS grouping may, in the future, provide an intuitive judgment of patients’ conditions through digital pathology, thus becoming an effective tool for precision treatment. At the same time, the assessment of
a
b
C
d
pvalue
pvalue
ESTIMATESOF
ESTIMATESCOPD
-É20
do
TOGA High 55
VAR GDO High 315
3.5
13
2
2
Z
2
TOGALOW SE
TOGA High SIS
I
3.0-
12-
5
Z
L
3
1
I
GEOLISIS
TOGA Low SIS
”
.
oboch
2.5.
11 -
Stomakicare
-
GEO High SIS
2
GEO Low SIS
.
2.0
10-
Turno Party
Z-score
1.5-
SIS
9-
SIS
-
- QUI Corelation Coefficient
1JI
-
Correlation Coefficient
tu
OS
absjoor]
1.0-
Low
8-
…
Score
11
Low
ESTIMATE SCONO
-6-01
0.5
High
7.
High
.
.
៛
2
6-
Iwrumešcomo
ESTIMATE SCONO
-40
ory calı
0.0-
Ssomalicomo
-0.5-
.
5
4-
0
-1.0-
TumorParty
inoParty
3.
-1.5
Correlation Coefficient
Correlation Coefficient
045
Cytotoxic Del
MHC
EC
SC
CP
IPS
e
f
g
Activity score
-4
0
low SIS
low SIS
L
Activity ROOM9
2
Step1 Step2
R=0.69, p = 2.20-12
Step3
I
10
Step4.T cell.recruiting
Step4.CD4
Step4.CD8 T cell.recruiting
-2.00
0
2.00
Stend mini coll.recruiting
Step4.Dendritic cell.recruiting
0
High SIS
Low SIS
Step4.Th22 cell.recruiting
Overall activity
Showerbeen recruiting
Stepd. Macrophage.fechoung
2
Step4. Neutrophilrecruiting Neutrophil.recruiting
1
*
Wound Healing
Step4.NK cell-recruiting
5
Macrophage Regulation
Step4. Eosinophil.recruiting
Step4.Basophil.recruiting
-20
Lymphocyte Infiltration
Step4. Th17 cell.recruiting
Step4.B cell.recruiting
3
IFN-gamma Response
TGF-beta Response
Step4. Th2 coll.recruiting
*
Step4.Treg cell.recruiting
.
-30
Step4.MDSC.recruiting
Step5
-5
D
SIS
5
10
h
68-06
100%
p =0.32
p =3.83e-05
100%
p =4.68e-03 p =0.02
p =0.41
p = 1.5De-07
p = 1.18e-03 p =2.18e-03
r
2
4.36-06
100% .
100%
90% -
90% -
90% -
16%
90% -
1
80% -
42%
80% -
30%
80% -
80% -
32%
70% -
1
70% -
56%
60% -
81%
SIS
70% -
60% -
72%
Responder
70% -
60% -
SIS
60% -
76%
Responder
TIDE
0
50% -
Low
50% -
TRUE
TIDE
50% -
High
50% -
TRUE
40% -
High
40%
D
70%
FALSE
40% -
84%
Low
40% -
68%
FALSE
30% -
58%
30%-
30% -
30% -
-1
20% -
20%
20% -
44%
20% -
10% -
19%
10% -
28%
-1
10% -
10% -
24%
0% -
(n = 36)
(n = 43)
0% -
(n = 50) (n = 29)
0% -
(n = 52)
(n = 58)
0% -
(n = 38) (n = 72)
High
Low
TRUE
FALSE
Low
High
-2
SIS
Responder
SIS
High
SIS
Low
TRUE
FALSE
High
Low
Responder
SIS
TCGA
GEO
j
k
log2FC(High vs Low)
-2
0
2
»
V
H
HLA-A
HLA-C
HIKE
TAP1
HLA-B
HLA-DMB
HLA-F
HLA-DOA
P
HLA-DOB
5
F
B2M
TAPZ
TAPBP
HLA-G
HLA-DPH
HLA-DRB
HLA-DQA
HLA-DOB
HLA-DELA
HLA-DRA
HLA-DRB1
CSF1R
CD160
KIR2DL
CLAVE
KIRZDL
A2aR
1001
B7-H4
IL1ORE LAG
LINGE
CD112
TGFBR
JEGFR
SLAMF
CTLA
CD96
DR.
PD-L
IL10
PD-L
TIME
LGALS9
TGFB1
CXCL12
CD48
B7.2
CD-0 VISTA
CO27
2025
STING
APRIL
CYCRA
MICA
RAETTE
CD40L ILBR
GIR
CO29
CD70
BINL
B7-15
ILBP
B7-H7
MICE
CD73
BAFF-R
OX40
CD30
B7.1
R7JE
CD267
ICOS
bump
COCO
NKG2A
TMIGD2 IMIGDZ
4- LIGHT
DR3
4-1BB-
HVEM
BAFF
OX-40L
Cancer_cell
ILE
LIA
CD155
Endothelial_cell
TCGA
Macrophage
Mesenchymal_cell
GEO
1
Progenitor_cell
ET LENE
Y
1
T_cell
MHC
Immunoinhibitors
Immunostimulators
I
m
n
Cortisol synthesis and secretion Steroid hormone biosynthesis
2.50
UBXN6
HOST
1.00
PI3K-Akt signaling pathway
1.50
CTRL
CD96
Pathways in cancer
Aldosterone synthesis and secretion
gFC(High vs Low)
GDF15
JACK
ITGA4
0.50
Proteoglycans in cancer
Steroid biosynthesis
Q.GD
0.60
Carbon metabolism
VCAM1
EPHAd
MAPK signaling pathway
-0.50
TMEM132A
0.40
Rap1 signaling pathway
Biosynthesis of amino acids Valine, leucine and isoleucine degradation
BMP4
GAB
RAMDO
Ras signaling pathway
-1.50-
GHR
RAMPS
0.20
EGFR tyrosine kinase inhibitor resistance
Pentose phosphate pathway
Propanoate metabolism
-2.50
FLT4
TUB83
PODXL
0.00
Thyroid hormone signaling pathway
Hedgehog signaling pathway
Glycine, serine and threonine metabolism
PTPRS
TOFERZ
Receptors
7AT1
Focal adhesion
Tryptophan metabolism
HSPA1ZA
EICH1
ECM-receptor interaction -
Cysteine and methionine metabolism
VGF
Cell adhesion molecules (CAMs)
Vitamin B6 metabolism
ATP4Α
W
COCO
Regulation of actin cytoskeleton
Sulfur relay system Sulfur metabolism
TAFA4
SORT1
Gap junction
mor exosom
CST5
AXL
ILTARA
Leukocyte transendothelial migration
Cell adhesion molecules (CAMs)
EPOR1
soPL
Cytokine-cytokine receptor interaction
Peroxisome
RAB22A
SALT
Natural killer cell mediated cytotoxicity
Arginine and proline metabolism
MRET
Jak-STAT signaling pathway
PRH1
Drug metabolism - cytochrome P450
PLAUS
PLAUR
NF-kappa B signaling pathway
Metabolism of xenobiotics by cytochrome P450
SLIT2
CCR1
T cell receptor signaling pathway
ABC transporters
TUB84A
SIGLEC10
Th17 cell differentiation
Terpenoid backbone biosynthesis
VTN
TGF-beta signaling pathway
Biosynthesis of unsaturated fatty acids
BAJAP2L1
TGFB1
Phagosome
FST
ANXA1
HIF-1 signaling pathway
0
3
6
9
ADGRV1
CD58
SEMA4D
0.0 2.5 5.0 7.5 10.012.5
Count
CHP1
CD48
Count
SERPINA3
HGF
Type
FN1
Ligands
Hormone Synthesis and Secretion Pathways
NEBL
THBS1
SEMA3B
ICAM1
Metabolic Reprogramming and Energy Metabolism Pathways
PDGFB
ASS1
CCL3
Type
Immune Evasion and Tumor Microenvironment Pathways
PFKP
SPP1
Tumor Pathways
Cell Growth and Signal Transduction Pathways
TOGA
GEO
Macrophage
T cat
Endothelial cell
Mesenchymal_call
Cancer_cell
Progenitor_cell
IL10
Macrophage
co8
Endothelial_coll
Mesenchymal_cell
Cancer Cell
Progenitor_cell
Cell Adhesion and Extracellular Matrix Pathways
Drug Metabolism and Resistance Pathways
Immune Pathways
Lipid Metabolism and Membrane Synthesis Pathways
Immune Evasion and Tumor Microenvironment Pathways
a
GDSC1
GDSC2
Rapamycin
BI-2536
BI-2536
CP466722
GW843682X
Ribociclib
P22077
Tipifarnib
Pyrimethamine
DMOG
GSK650394
WZ-1-84
Mitomycin-C
Doxorubicin
Epothilone B
Gemcitabine
Vinorelbine
Mitoxantrone
Topotecan
BMS-754807
BMS-754807
PRT062607
Entospletinib
NVP-ADW742
GNF-2
Thapsigargin
Sepantronium bromide
AZD5991
AZD1208
Enzastaurin
QS11
MIM1
Tozasertib
AS605240
Idelalisib
ZSTK474
AZD6482
JQ1
UMI-77
AZD8055
SB216763
Doramapimod
WZ4003
Afuresertib
PI3K/Akt/mTOR
Autophagy
Apoptosis
Chemotherapy
Cell Cycle/DNA Damage Epigenetics
Protein Tyrosine Kinase/RTK
Metabolic Enzyme/Protease
Others
Membrane Transporter/Ion Channel
TGF-beta/Smad
Immunology/Inflammation
JAK/STAT Signaling
MAPK/ERK Pathway
TGF-beta/Smad
Stem Cell/Wnt
mTOR
PLK1
PLK1
ATM
PLK1&PLK3
CDK4/6
USP7
HIF-PH
SGK
FTase
ARFGAP1
IGF-1R/IR
IGF-1R/IR
Syk
Syk
Bcl-Abl
IGF-1R
Ca2+-ATPase
Mcl-1
Mcl-1
PIM
survivin
PKCß
BET bromodomain
Aurora A/B/C
PI3KY
p1105
PI3K
p110ß
Mcl-1
mTOR
GSK-3
p38 MAPK
NUAK kinase
pan-Akt
b
C
cd
e
SPIED3 correl
cMAP Score
0
1
-100
0
100
RIBONUCLEOTIDE_BINDING
GEO
TCGA
GEO
TOGA
cMAP
SPIED3
ADENYL_NUCLEOTIDE_BINDING-
Count
Acarlylehalina receptor antagonist
ZK-93428
Benzodiakhaipine receptor antagonist
diflorinone
Corticosteroid agonist
10
salbutamol
Adrenergie mecuptor agonist
triploliche
RNA polymerinie inhibilor
15
lelodipina
Antiviral
purmorphimine
Smoothanad niciplor iagonisit
20
apigunin
Calcium channel blocker
Cassin kinicia inhibitor
diazep
Acetylcholine receptor antagonist
OXIDOREDUCTASE_ACTIVITY -
O
Adenosine reuptake inhibitor
25
thiocolchicmida
Adninergie receptor antagonist
etholpin upagliniche
nicargoline
Adrinergie receptor antagonist
30
diodromethe
Hydantoin ansapikipdie
BCR-ABL kinase inhibitor
35
larryleypromina
Insulin secretagogue
CALCIUM_ION_BINDING-
Monoamine dedise inhibitor
Calmedulin antagonist
OF handling respeto
:40
H3-504393
CC chemokine nicaptor antagonist
LY-294002
MTOR inhibitor
promicine
Dopamine nacaptor antagonist
dipyridamal
pressinbinding protein inhibitor
R
BIBX-1382
depropor antagonist
SULFUR_COMPOUND_BINDING-
-log10(pvalue)
O
EGFR inhibitor
BIBU-1361
TG-101348
EGFR inhibitor
FLT3 inhibitor
1.4
daunorubin
ellipticine
RINA synthesis inhibitor
nebúverina
O
Topoisomine
AY-0944
Hedgehog pathway modulator Histamine receptor agonist
1.6
Acıılylcholine receptor antagonist
alimentprint
Adrenergie receptor antagonist
BIX-01294
Histone lysine methyltransferase inhibitor
1.8
lapinesib
Kimarin-like spindle protiin inhibitor
ENZYME_INHIBITOR_ACTIVITY -
zbavitin
Antiviral
2.0
Aromalinie inhibitor
ML-7
Neural Wiikoll-Aldrich syndrome protein inhibitor
Acetylcholine receptor antagonist Dopamine receptor antagonist BCR-ABL kinase inhibitor Norepinephrine reuptake inhibitor
2.2
woikostatin
erythromycin
NFKB pathway inhibitor
2.4
prochlorpanazina
maprotiline
donapine
Dopamine receptor antagonist
TETRAPYRROLE_BINDING
2.6
FIT
Nonpinephrine reuptake inhibitor
U-18868A
Opioid receptor agonist
2.8
tuciopunthinol
Dopamine nicaptor antagonist
Dopamina nicaptor
L-388899
Oxidoriqualene cyclase inhibitor
CARA riktigis inhibitor
Cx0562 inhibitor
H-80
Oxytocin receptor antagonist PKA.inhibitor
Sucocorchia receptor agonist Holamcal nicuptor antagonist
RS-39804
Serotonin receptor antagonist
R-96544
Serotonin receptor antagonist
IRON_ION_BINDING-
HMOCR inhibitor
GR-55562
Burtonin Nowparaganar
rosiglitazone
lesulin seraitiper
pentaxilylling
Nonspinaphrine reuptake inhibitor
S8-216841
Serotonin receptor anbigonist.
0.05
Phosphodiesterase inhibitor
Stol damethylase inhibitor
0.10
0.15
0.20
quipazine
Serotonin receptor agonist
Viraicular mondumine trariporter inhibitor
GeneRatio
f
Iron metabolism & ferroptosis
Calcium metabolism
Schizophrenia
Macrophage
1.00
T_cell
0.80
Endothelial_cell
0.60
Mesenchymal_cell
0.40
Cancer_cell
0.20
0.00
AKR1C3
FADS? NR1D
CHO.
CHP1
AOX1
PEY?
ALDHS
11A
CYP112
CYP2145
OLD
ISCA2
GDF15
BEX1
MMP 16 CCBE1
RGN
CALB:
NECAB3
CHP:
COM11
WNK3
CACNA1H
WOL3
CD320
EPDR1
FAM155A
SPA SLITS
Progenitor_cell
CEIN2
CALN1
CACNA 1D
TUBBA
SPOCK 1
PCI
ADGRL3
ANO4
REPS2
ROBO1
OLG2
RBFOX1
CNNM2
NOVA
ANKS1B
SAMP
AS3MT
MCLO
ANY
TUBB3
NPAS3
MAOA
GRING
GRID2
SERPINA3
lymphocytic infiltration may also provide an essential complement to the Weiss scoring system.
Our analysis highlights the key role of immune infiltration and ster- oidogenic pathways in ACC treatment. CD8 + T cells and Tfh cells play pivotal roles in anti-tumor immunity27, exhibiting greater levels of immune
infiltration within the high SIS group. Moreover, more than half of the patients in this group showed no abnormal hormone secretion, further underscoring the association of high SIS with improved prognosis. TIDE analysis revealed that over 70% of individuals with high SIS showed a favorable response to ICB, indicating that immunotherapy, particularly
ICB, could be more beneficial for this subgroup. Conversely, those in the low SIS category exhibited reduced overall immune infiltration but elevated levels of M2 macrophages, which are known to promote tumor progression and impair T-cell function28, potentially contributing to the limited effec- tiveness of immunotherapy in these patients. Using single-cell RNA sequencing, we have gained insight into how tumor cells affect immune cell communication, particularly the ligand-receptor pairs of T cells and mac- rophages, including the known ligand-receptor pairs of CCL3-CCR1, PDGFB-PDGFRB, TGFB1-TGFBR2, VCAM1-ITGA4, and FN1-ITGA4. Among them, the binding of tumor-secreted VCAM1 to T cell surface receptor ITGA4 may prevent T cells from adhering to tumor cells and limit their attack; meanwhile, VCAM1 competitively binds ITGA4 to FN1, which further inhibits T cell function, revealing a novel mechanism of immune escape from ACC. EGFR is overexpressed in the majority of ACC patients29, however, EGFR tyrosine kinase inhibitors, such as erlotinib and gefitinib, demonstrate limited efficacy30. Our single-cell RNA sequencing analysis identified ligand receptors (e.g., AXL, HGF, PDGFB, PDGFRB) in macro- phages and mesenchymal stromal cells as potential contributors to drug resistance, offering fresh insights into the mechanisms underlying drug resistance in ACC.
Given the substantial number of patients in the low SIS group and their poor prognosis-over 75% of whom exhibit abnormal hormone secretion-prioritizing the use of hormone-suppressing drugs is essential. Mitotane is the only FDA-approved drug that inhibits corticosteroid synthesis, and despite more than 50 years of clinical use, its exact mechanism remains unclear. Current evidence suggests that it limits steroid production by inhibiting the activity of steroidogenic enzymes, thereby preventing the conversion of cholesterol to steroids31. In our study, machine learning identified the steroid synthesis gene most closely associated with the SIS signature, DHCR7. DHCR7 catalyzes the final step of cholesterol synthesis by converting 7-dehydrocholesterol into cholesterol32, and serves as an upstream gene of Mitotane’s known ther- apeutic target. In pan-cancer, DHCR7 expression is highest in ACC and is linked to unfavorable survival outcomes. Knockdown of DHCR7 increases the sensitivity of ACC to Mitotane, suggesting that DHCR7 may be a novel therapeutic target.
Our study suggests that patients with high SIS are likely to respond better to drugs targeting the PI3K/Akt/mTOR pathway, whereas those with low SIS may show greater sensitivity to alternative therapies. In addition to the PLK1 inhibitors predicted by GDSC1 and GDSC2, the CMap database suggests that various drugs capable of inhibiting steroid hormone synthesis may be effective in the low SIS group, including clozapine. As a treatment for schizophrenia, clozapine has been shown to inhibit aldosterone secretion through inhibition of the D4 receptor20. In addition, our study found that schizophrenia is associated with survival-ImmunQTLs in multiple immune cells in ACC. Single-cell RNA sequencing showed that genes specifically expressed in ACC tumor cells were enriched in iron and calcium metabo- lism pathways, which are highly associated with schizophrenia. Therefore, we suggested that medications for schizophrenia might have the potential to be co-administered with mitotane based on our analysis. On the other hand, drugs that inhibit calcium metabolism may be effective in all patients within the SIS group. Notably, Ca2+ levels in ACC immune cells were most strongly associated with the Survival-ImmunQTLs. Therefore, drugs inhi- biting calcium metabolism and promoting ferroptosis may have important research value in tumor immunity and progression in ACC.
This study not only deeply characterizes the immune micro- environment of ACC but also reveals the important roles of metabolic and immune-synergistic regulatory gene clusters in immune and hormonal therapy response in ACC patients, providing new perspectives for exploring the mechanisms of ACC genesis and also laying the foundation for future personalized treatment strategies that may change the way we diagnose and treat this cancer. This study advances the integration of digital pathology and AI in oncology, facilitates the implementation of precision medicine strategies, and leads to innovative solutions for treating rare diseases.
Methods
Data collection on ACC
We obtained gene expression profiles and clinical information on ACC from the TCGA portal (http://cancergenome.nih.gov) and used 79 ACC samples for subsequent analysis. Among them, WSIs from 55 patients (a total of 237 slides) were sourced from the TCGA database. Four ACC datasets with prognostic information were downloaded from GEO, which are GSE33371, GSE10927, GSE76019, and GSE76021. These were merged to remove the batch effect using the function “ComBat” to remove batch effects33.
Calculation of cell abundance in the immune microenvironment
Using the R package “GSEAbase,” we examined 24 microenvironmental cell subpopulation-related expression levels in the TCGA, GSE76019, and GSE76021 datasets for single-sample gene set enrichment analysis (ssGSEA) concerning the immune cell profiles constructed in the published article34.
Differentially expressed genes (DEGs) of TIME clusters
The R package “limma” identified common DEGs in TIME clusters. DEGs with FDR values < 0.001 and absolute fold change > 1 were considered significant and used for further analysis.
Consensus clustering of DEGs
To deeply reveal the immunological significance of ACC, we used the expression profiling data of DEGs to perform consensus clustering analysis using the “ConsensusClusterPlus” package. The optimal classification (k=2) was determined by calculating the consensus matrix and the con- sensus cumulative distribution function.
Development of the Steroid-related Immune Score (SIS) System
First, unsupervised clustering was performed based on differentially expressed genes (DEGs) to classify ACC patients into two gene subtypes (Gene Cluster A and B). Subsequently, based on the characteristics of the gene clusters, 18 DEGs were further categorized into SIS gene signature A (genes positively correlated with cluster characteristics) and SIS gene sig- nature B (genes negatively correlated with cluster characteristics). To enhance model stability and interpretability, the Boruta algorithm was applied for dimensionality reduction of SIS gene signatures A and B. Principal component analysis (PCA) was then performed to extract the first principal component (PC1) as the final scoring criterion, where PC1A represents the principal component score for SIS gene signature A, and PC1B represents the principal component score for SIS gene signature B. Finally, the SIS score was calculated by aggregating PC1A and PC1B (SIS = EPCIA + EPC1B), a method similar to the Gene Expression Grade Index (GGI). Additionally, we provide the PC1 values and their contribution weights involved in the SIS calculation to enhance the transparency and reproducibility of our study (Supplementary Table 1-3).
Image preprocessing, data partitioning, and external validation
Image preprocessing included patch extraction, color normalization, and filtering of staining artifacts to ensure the quality and consistency of model input data. First, WSIs were divided into 256 x 256-pixel patches at 10x magnification using the OpenSlide library, and patches with more than 50% background (RGB color values below 220) were discarded. Next, color normalization was performed using the Macenko method, and additional patches with staining artifacts were filtered out to minimize the impact of technical variations35. After preprocessing, a total of 931,162 patches were generated, with the number of patches per WSI ranging from 544 to 7609.
The dataset consisted of 55 ACC patients from TCGA (20 in the high- SIS group and 35 in the low-SIS group). To enhance the robustness and generalizability of the model, five-fold cross-validation was employed. Data partitioning was conducted at the patient level, ensuring that each fold contained 4 high-SIS patients and 7 low-SIS patients. In each training iteration, the dataset was split into a training set, validation set, and test set in
a 3:1:1 ratio, and the final results were reported as the average performance across five test runs. During training, all slides and patches from each patient inherited the corresponding patient-level label, and classification models were trained at the patch level. Given that each patient had multiple slides, with each slide containing thousands of patches, a resampling strategy was applied to balance the dataset and mitigate class imbalance. Approximately 24,000 patches were used per fold during training.
To further evaluate the generalizability of the model, we included an independent external validation cohort comprising 20 ACC patients from the First Affiliated Hospital and the Second Affiliated Hospital of Dalian Medical University. This dataset underwent the same preprocessing pipe- line and was used for inference with the trained model to assess its classi- fication performance in an independent cohort.
Integration of model prediction results, transfer learning, and evaluation metrics
During the testing phase, the model predicts all patches of all cases in the test set and calculates the predicted category and confidence score for each patch. For all patches within the same slide, we apply a confidence-weighted average to determine the final classification result of that slide. Furthermore, the final classification result of each patient is determined by the weighted average of all their slides, thereby enhancing the model’s stability and reliability at the patient level.
To mitigate the risk of overfitting due to small-sample deep learning training, we employed transfer learning during model training. Specifically, we utilized ResNet50 for transfer learning, fine-tuning it based on a pre- trained model from ImageNet to leverage existing feature representations and improve model generalization. Additionally, we explored the Vision Transformer (ViT-B16) architecture based on the self-attention mechanism to further evaluate the impact of different model architectures on the classification task.
The classification performance of the model was assessed using five- fold cross-validation and quantified by multiple evaluation metrics, including Accuracy, AUC (area under the receiver operating characteristic curve), Recall, and Specificity. All results were calculated as the average across the five-fold cross-validation test sets and further evaluated on an external validation cohort to ensure the robustness and generalizability of the classification model.
Patient Selection and Clinical Sample Collection
We retrospectively identified 10 ACC patients who underwent surgical procedures in the Department of Urology at the Second Affiliated Hospital of Dalian Medical University from January 1, 2012, to March 1, 2024, and 10 ACC patients who underwent surgery in the Department of Urology at the First Affiliated Hospital of Dalian Medical University during the same period. We obtained data on hematoxylin and eosin (H&E) sections and relevant clinical information for these patients. We digitized the H&E sections using a panoramic digital pathology slide scanner (BL-006, Songming Medical Tech) at 40x (0.25 um/pixel) magnification. We then performed image pre-processing, data segmentation, model prediction, and validation using the abovementioned methods. We collected a fresh spe- cimen of ACC during surgical resection for subsequent single-cell RNA sequencing. We conducted the study under the Declaration of Helsinki. The Ethics Committee of the Second Affiliated Hospital of Dalian Medical University (No. KY2024-032-01) and the Ethics Committee of the First Affiliated Hospital of Dalian Medical University (No. PJ-KS-KY-2025-27) approved all tissue samples included in this study. Each patient provided written informed consent.
Single-cell RNA sequencing(scRNA-seq) and related analysis
A fresh ACC specimen was collected from the Second Affiliated Hospital of Dalian Medical University and transported to the laboratory on ice in MACS tissue storage solution (Miltenyi Biotec). Once the sample was made into a single-cell suspension, we used a Countstar Fluorescence Cell Ana- lyzer (Countstar) to measure cell counting and viability, adjusting the cell
concentration to 300 — 600 cells/uL. In accordance with the manufacturer’s protocol, we introduced the cell suspension into a 10 x Genomics Chro- mium Controller to produce a single-cell gel bead emulsion. We used the single-cell 3’ Library and Gel Bead Kit V3.1 (10x Genomics, 1000121) to make single-cell RNA-seq libraries, and an Illumina Novaseq6000 was used for sequencing. The data were then processed using CellRanger (version 3.0.2), and the gene expression matrix was imported into Seurat (version 3.0) for quality control and downstream analysis. In Seurat 3.0, cells with more than 200 genes and ≤25% mitochondrial gene expression were selected and retained. PCA was used for dimension reduction, and t-SNE was used for visualization. We selected the top five genes, ranked from largest to smallest in avg_log2FC, as candidate marker genes for this cell subset, according to the conditions avgUMI≥1 & p_val_adj≤0.05. We selected genes specifically expressed in cancer cells (avg_log2FC>0.4) and queried their ability to become exosomes from Genecards. We obtained the ligands and receptors of immune cells from CellphoneDB (v5.0.0) (https:// www.cellphonedb.org/)36. HitPredict (https://www.hitpredict.org/#/) is an experimentally validated protein interaction website37. We manually quer- ied the interaction between tumor exosomes and immune cell ligands and receptors and plotted a heatmap.
Immunohistochemistry (IHC) of ACC
Dewax the sections in the following order: xylene, absolute ethanol, and ethanol of different concentrations. Then, rinse with water. The antigen retrieval was performed using 1x EDTA (pH 8.0) repair solution in a microwave, and the solution was allowed to return to room temperature after heating. Incubate the sections at room temperature for 30 min in 3% H2O2 and then wash with PBS. Circle the tissue with a histochemical pen, add 3% BSA dropwise, and incubate at 37 ℃ for 30 min to perform serum blocking. After blocking, remove the serum and add the anti-DHCR7 rabbit polyclonal antibody (1:200, ZEN-BIOSCIENCE, 822232) dropwise to the tissue. Incubate overnight at 4℃. Next, incubate the tissue at 37 °℃ for 1 h with horseradish peroxidase (HRP)-labeled anti-rabbit secondary antibody (JiJia Biotechnology, J0046) and then wash three times with PBS. Add the DAB working solution dropwise to develop color. Observe under a microscope for specific brown expression. Then, rinse with water. After staining the slide with hematoxylin, differentiate it with hydrochloric acid alcohol, counterstain with blueing solution, and dehydrate with absolute ethanol and n-butanol; add neutral gum to seal the slide and let it dry. The DHCR7 immunohistochemistry of normal adrenal tissue is selected from THE HUMAN PROTEIN ATLAS (https://www.proteinatlas.org/)38.
Calculation of common tumor markers and access to genomic alterations information
From the databases dbSNP and ExAC, we filtered all germline mutations. After that, we computed and defined the TMB of every sample using the total number of coding variants/length of exons (38 million). Variants were regarded as missense mutations, in-frame deletions, in-frame insertions, frameshift deletions, frameshift insertions, nonsense mutations, stop mutations and silent mutations. Based on the median TMB value, we split patients into low TMB group and high TMB group. We evaluated the EMT score of ACC patients using the EMT gene signature constructed in a published article as a ref. 39. Based on the median EMT value, patients were split into low and high EMT groups. We calculated the mRNAsi of ACC using the OCLR machine learning algorithm40. Median mRNAsi value guided patients’ division into high and low mRNAsi groups. We obtained genomic alteration information from cbioportal (https://www.cbioportal. org/)41 and selected genes with p <0.05 for comparison according to SIS grouping (two-sided Fisher Exact test).
Gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA)
We used the R package “GSVA” to comprehensively score the gene sets and perform differential analysis between subgroups. We used the GSEA
software (version 4.3.0) to analyze gene set enrichment between subgroups. The gene sets were obtained from the Molecular Signatures Database (h.all.v2022.1.Hs and c2.cp.kegg.v2022.1.Hs).
Cell culture, transfection and pharmacological assays
Procell Life Science & Technology provided human ACC cell lines (SW13, NCI-H296R). Ten percent FBS (PAN Seratech, ST30-3302) and one percent P/S (Procell, PB180120) were included into DMEM media (EallBio) for SW13 cells. Ten percent FBS, one percent P/S, and 0.5% insulin-transferrin- selenium supplement (ITS-G) (Procell, PB1804) were included in DMEM/ F12 media (EallBio, 03.1012) used for NCI-H295R cells. Once the cells reached 60-80% confluence, they were digested and collected. We counted the cells and seeded them into 96-well plates. Once the cells had been attached, the control wells were treated with different concentrations of mitotane, and an equal volume of DMSO was added. Each well received 100 µL of 10% CCK-8 solution (ApexBio, K1018) following 48 h of treatment; this was added, incubated at 37 ℃ for 2 hours, and the absorbance was measured at 450 nm.
The day before transfection, seed the cells in a 6-well plate and allow them to reach 40-60% confluence. Mix the siRNA with Lipofectamine™ 2000 (Life Technologies) as instructed by the manufacturer. Add the mixture to the well plate and incubate for six hours. Swap the medium for a complete one with 10% FBS. The cells were gathered 48 hours into culture, and qRT- PCR verified the expression level of DHCR7 to confirm the knockdown effect. We counted the cells after transfection and seeded them into a 96-well plate. We then performed drug treatment using the IC50 concentration we had measured above. Each well received 100 µL of 10% CCK-8 solution (ApexBio, K1018) following 48 hours of treatment; this was added, incu- bated at 37 ℃ for 2 h, and the absorbance was measured at 450 nm.
Immune-related indicators
Based on the features connected with stromal tissue and immune cell infil- tration, the ESTIMATE algorithm defines the stromal and immune score of every patient12. The ESTIMATE score is the total of the stromal and immune values. We scored each ACC sample using the R package “estimate.” To ensure data consistency, we selected the Tumor Immune MicroEnvironment cell composition Database (TIMEDB) (https://timedb.deepomics.org/) to obtain the relative abundance of immune cells in ACC patients using different algorithms (CIBERSORT, ABIS, MCPcounter, xCell, and ImmuCellAI)43. We analyzed the anti-cancer immune status in the ACC immune cycle using TIP (http://biocc.hrbmu.edu.cn/TIP/)44. We obtained the ACC immunopheno- score (IPS) from The Cancer Immunome Atlas (https://tcia.at/), which includes MHC molecules (MHC), Checkpoints (CP), effector cells (EC, such as activated CD8 + T cells and CD4 + T cells, Tem CD8+ and Tem CD4+ cells), and suppressor cells (SC, e.g., Tregs and MDSCs)13. In a published article, we assessed the differences between ACC SIS subgroups concerning the five identified representative signatures of pan-cancer immunity14.
To forecast each ACC tumor sample’s response to immune checkpoint blockade, we also computed the TIDE score for each utilizing TIDE (http:// tide.dfci.harvard.edu/faq/)15. We also investigated the impact of genetic variation on immune infiltration in ACC utilizing CancerImmunityQTL (http://www.cancerimmunityqtl-hust.com/)45. Finally, we generated DNA damage scores using ABSOLUTE by TCGA aneuploidy AWG46,47.
Prediction of SIS-related sensitive drugs
The Genomics of Drug Sensitivity in Cancer (GDSC; https://www. cancerrxgene.org) database is the definitive source for assessing cancer cells’ drug sensitivity and response48. We used the “oncoPredict” package to determine the drug sensitivity estimates of ACC patients to GDSC1 and GDSC2 drugs49. We then selected the drugs that differed between the SIS subgroups (p<0.05). We input the differentially expressed genes (|logFC |>1) of the SIS subgroups into CMap (https://clue.io/) and SPIED3 (http://92.205.225.222/HGNC-SPIED3-QF.py)50,51. We selected the drugs with a |Score | >90 in CMap and the top 100 positively and negatively correlated drugs in SPIED3.
Statistical analysis
R 4.2.1 software was used in all statistical analyses. We did differential analyses with the limma package. We used the student’s t test to examine continuous variables and the x2 test to examine categorical variables. Comparisons between two groups and between two or more groups were conducted using the nonparametric two-sided Wilcoxon rank-sum test and the Kruskal-Wallis test, respectively. We computed correlations using Spearman’s coefficient. The log-rank test helped us to determine the statistical relevance of survival rates between different subgroups. To investigate significant prognostic elements, we conducted multivariate and univariate Cox regression analysis.
Data availability
All data supporting the findings of this study will be made available upon reasonable request.
Received: 28 November 2024; Accepted: 11 August 2025; Published online: 18 September 2025
References
1. Kerkhofs, T. M. et al. Adrenocortical carcinoma: a population-based study on incidence and survival in the netherlands since 1993. Eur. J. Cancer 49, 2579-2586 (2013).
2. Fassnacht, M. et al. Limited prognostic value of the 2004 international union against cancer staging classification for adrenocortical carcinoma: proposal for a revised tnm classification. Cancer 115, 243-250 (2009).
3. Stigliano, A. et al. Management of adrenocortical carcinoma: a consensus statement of the italian society of endocrinology (sie). J. Endocrinol. Invest. 39, 103-121 (2016).
4. Amini, N. et al. Curative resection of adrenocortical carcinoma: rates and patterns of postoperative recurrence. Ann. Surg. Oncol. 23, 126-133 (2016).
5. Fassnacht, M., Kroiss, M. & Allolio, B. Update in adrenocortical carcinoma. J. Clin. Endocrinol. Metab. 98, 4551-4564 (2013).
6. Puglisi, S. et al. New perspectives for mitotane treatment of adrenocortical carcinoma. Best. Pract. Res. Clin. Endoc. Metab. 34, 101415 (2020).
7. Baxi, V., Edwards, R., Montalto, M. & Saha, S. Digital pathology and artificial intelligence in translational medicine and clinical practice. Mod. Pathol. 35, 23-32 (2022).
8. Ghosh, C., Hu, J. & Kebebew, E. Advances in translational research of the rare cancer type adrenocortical carcinoma. Nat. Rev. Cancer 23, 805-824 (2023).
9. Coudray, N. et al. Classification and mutation prediction from non- small cell lung cancer histopathology images using deep learning. Nat. Med. 24, 1559-1567 (2018).
10. Kather, J. N. et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat. Med. 25, 1054-1056 (2019).
11. Zheng, S. et al. Comprehensive pan-genomic characterization of adrenocortical carcinoma. Cancer Cell 29, 723-736 (2016).
12. Assie, G. et al. Integrated genomic characterization of adrenocortical carcinoma. Nat. Genet. 46, 607-612 (2014).
13. Charoentong, P. et al. Pan-cancer immunogenomic analyses reveal genotype-immunophenotype relationships and predictors of response to checkpoint blockade. Cell Rep. 18, 248-262 (2017).
14. Thorsson, V. et al. The immune landscape of cancer. Immunity 48, 812-830 (2018).
15. Jiang, P. et al. Signatures of t cell dysfunction and exclusion predict cancer immunotherapy response. Nat. Med. 24, 1550-1558 (2018).
16. Joyce, J. A. & Fearon, D. T. T cell exclusion, immune privilege, and the tumor microenvironment. Science 348, 74-80 (2015).
17. Ayers, M. et al. Ifn-gamma-related mRNA profile predicts clinical response to PD-1 blockade. J. Clin. Invest. 127, 2930-2940 (2017).
18. Spat, A., Hunyady, L. & Szanda, G. Signaling interactions in the adrenal cortex. Front. Endocrinol. 7, 17 (2016).
19. Hall, P. F. The roles of calmodulin, actin, and vimentin in steroid synthesis by adrenal cells. Steroids 62, 185-189 (1997).
20. Wu, K. D. et al. Expression and localization of human dopamine d2 and d4 receptor mma in the adrenal gland, aldosterone-producing adenoma, and pheochromocytoma. J. Clin. Endocrinol. Metab. 86, 4460-4467 (2001).
21. Sh, S. & Sm, H. Should oxidosqualene cyclase in the cholesterol biosynthetic pathway be considered an anti-cancer target? Front. Cell. Dev. Biol. 10, 1081151 (2022).
22. Lefebvre, H., Contesse, V., Delarue, C., Vaudry, H. & Kuhn, J. M. Serotonergic regulation of adrenocortical function. Horm. Metab. Res. 30, 398-403 (1998).
23. Zarn, J. A., Bruschweiler, B. J. & Schlatter, J. R. Azole fungicides affect mammalian steroidogenesis by inhibiting sterol 14 alpha-demethylase and aromatase. Environ. Health Perspect. 111, 255-261 (2003).
24. Identification of risk loci with shared effects on five major psychiatric disorders: a genome-wide analysis. Lancet 381, 1371-1379 (2013).
25. Lotan, A. et al. Perturbed iron biology in the prefrontal cortex of people with schizophrenia. Mol. Psychiatry 28, 2058-2070 (2023).
26. Liu, D. et al. Schizophrenia risk conferred by rare protein-truncating variants is conserved across diverse human populations. Nat. Genet. 55, 369-376 (2023).
27. Crotty, S. T follicular helper cell biology: a decade of discovery and diseases. Immunity 50, 1132-1148 (2019).
28. Gao, J., Liang, Y. & Wang, L. Shaping polarization of tumor- associated macrophages in cancer immunotherapy. Front. Immunol. 13, 888713 (2022).
29. Adam, P. et al. Epidermal growth factor receptor in adrenocortical tumors: analysis of gene sequence, protein expression and correlation with clinical outcome. Mod. Pathol. 23, 1596-1604 (2010).
30. Konda, B. & Kirschner, L. S. Novel targeted therapies in adrenocortical carcinoma. Curr. Opin. Endocrinol. Diab. Obes. 23, 233-241 (2016).
31. Corso, C. R. et al. Pharmacological profile and effects of mitotane in adrenocortical carcinoma. Br. J. Clin. Pharmacol. 87, 2698-2710 (2021).
32. Prabhu, A. V., Luu, W., Li, D., Sharpe, L. J. & Brown, A. J. Dhcr7: a vital enzyme switch between cholesterol and vitamin d production. Prog. Lipid Res. 64, 138-151 (2016).
33. Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical bayes methods. Biostatistics 8, 118-127 (2007).
34. Xiao, Y. et al. Multi-omics profiling reveals distinct microenvironment characterization and suggests immune escape mechanisms of triple- negative breast cancer. Clin. Cancer Res. 25, 5002-5014 (2019).
35. M. et al. A method for normalizing histology slides for quantitative analysis. In Proc. IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 1107-1110 (IEEE, 2009).
36. Garcia-Alonso, L. et al. Single-cell roadmap of human gonadal development. Nature 607, 540-547 (2022).
37. Lopez, Y., Nakai, K. & Patil, A. Hitpredict version 4: comprehensive reliability scoring of physical protein-protein interactions from more than 100 species. Database 2015 (2015).
38. Uhlen, M. et al. Proteomics. Tissue-based map of the human proteome. Science 347, 1260419 (2015).
39. Gibbons, D. L. & Creighton, C. J. Pan-cancer survey of epithelial- mesenchymal transition markers across the cancer genome atlas. Dev. Dyn. 247, 555-564 (2018).
40. Malta, T. M. et al. Machine learning identifies stemness features associated with oncogenic dedifferentiation. Cell 173, 338-354 (2018).
41. Cerami, E. et al. The cbio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov. 2, 401-404 (2012).
42. Yoshihara, K. et al. Inferring tumour purity and stromal and immune cell admixture from expression data. Nat. Commun. 4, 2612 (2013).
43. Wang, X. et al. Timedb: tumor immune micro-environment cell composition database with automatic analysis and interactive visualization. Nucleic Acids Res. 51, D1417-D1424 (2023).
44. Xu, L. et al. Tip: a web server for resolving tumor immunophenotype profiling. Cancer Res 78, 6575-6580 (2018).
45. Tian, J. et al. Cancerimmunityqtl: a database to systematically evaluate the impact of genetic variants on immune infiltration in human cancer. Nucleic Acids Res. 49, D1065-D1073 (2021).
46. Carter, S. L. et al. Absolute quantification of somatic dna alterations in human cancer. Nat. Biotechnol. 30, 413-421 (2012).
47. Taylor, A. M. et al. Genomic and functional approaches to understanding cancer aneuploidy. Cancer Cell 33, 676-689 (2018).
48. Yang, W. et al. Genomics of drug sensitivity in cancer (gdsc): a resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Res. 41, D955-D961 (2013).
49. Maeser, D., Gruener, R. F. & Huang, R. S. Oncopredict: an r package for predicting in vivo or cancer patient drug response and biomarkers from cell line screening data. Brief. Bioinform. 22, bbab260 (2021).
50. Subramanian, A. et al. A next generation connectivity map: l1000 platform and the first 1,000,000 profiles. Cell 171, 1437-1452 (2017).
51. Williams, G. Spiedw: a searchable platform-independent expression database web tool. Bmc Genomic. 14, 765 (2013).
Acknowledgements
This study was supported by the following grants: the Scientific Research Project of the Ministry of Education of Liaoning Province (Grant No. LJKZZ20220100); the Interdisciplinary Research Cooperation Project Team Funding of Dalian Medical University, Planning and Research Category (focusing on planning for recreation) (Grant No. JCHZ2023001); and the Joint Foundation of the Dalian Institute of Chemical Physics, Chinese Academy of Sciences, and the Second Hospital of Dalian Medical University (Grant No. DMU-2 & DICP UN202304).
Author contributions
Z.L. designed and supervised the study, secured funding, and provided critical revisions to the manuscript. D.Y. and H.W. were responsible for the overall direction and planning of the study. W.H. oversaw the study’s design, performed most of the experiments, carried out data analysis, and prepared the initial manuscript draft. L.Y. and J.W. contributed to the study design and were responsible for AI analysis, co-writing significant portions of the manuscript. Y.Z. supported experimental work, managed data curation, and contributed to manuscript revisions. Z.D. led the single-cell sequencing analysis and provided critical feedback on data interpretation. X.S. con- ducted analyses to support AI-driven pathology recognition. B.F. and Y.W. provided technical support for specific experimental techniques and data processing. H.X. and Y.W. collected clinical data, including digital conver- sion of patient H&E slides, from the First and Second Affiliated Hospitals of Dalian Medical University, respectively. X.G., P.L., H.Z., L.W. and Y.W. also provided technical support for specific experimental techniques and data processing. All authors reviewed, revised, and approved the final version of the manuscript.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41698-025-01092-4.
Correspondence and requests for materials should be addressed to Hongyu Wang, Deyong Yang or Zhiyu Liu.
Reprints and permissions information is available at http://www.nature.com/reprints
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by- nc-nd/4.0/.
C The Author(s) 2025